A Gyroscope Error Separation-Based Method for Alignment and Stitching of Noisy Videos

In low-light environments, the alignment and stitching of noisy videos can not only broaden the field of view but also facilitate image denoising. However, image-based alignment methods typically produce poor results and cannot meet dynamic requirements in this case. In this paper, we propose a new method for high-accuracy alignment and stitching of frames in noisy videos. By analyzing the characteristic of gyroscope errors in the time domain, we propose an online correction model. First, we novel separate gyroscope errors into deterministic and random components through time-domain signal decomposition. Second, we roughly approximate the deterministic component by combining gyroscope data with image matching techniques. Finally, we precisely calculate the deterministic component in the time domain and compensate it to gyroscope data, resulting in a precise rotation vector for image alignment and stitching. Utilizing a two-stage methodology allows for precise approximation of deterministic components, thereby significantly minimizing the estimation error inherent to deterministic components. Experiments show that our method achieves high accuracy and strong robustness compared to other methods through testing on noisy videos.